Ask an Analyst: Remote Sensing for Agricultural Programming in Sudan
At the end of 2023, we wrote about how Mercy Corps was leveraging geospatial analysis to assess the impacts of the ongoing war in Sudan. With nearly a year gone by since the start of the conflict, our country team has once again found itself thinking about agricultural assistance for the next growing season. To that end, Mercy Corps’ T4D, Global Crisis Analysis, and the Sudan Country Team teamed up with Nick Dowhaniuk, a geospatial data scientist, and Sarah McArthur, a conflict analyst, to develop projections of cropland vegetation health through July 2024.
Advancing our work from last year, this analysis used satellite imagery and machine learning to construct a model to spatially predict vegetation health based on expected weather conditions, with secondary data analysis to understand the interaction of weather conditions with other drivers of agricultural productivity. Layered with contextual interpretation, the analysis provides an overview of risks and opportunities in the agricultural sector in four states of Sudan: Gedaref, Blue Nile, South Kordofan, and Kassala. You can read the reports here.
We sat down and discussed the modeling work and its outputs with Nick.
To start out, can you share a bit about what the guiding research questions were as you developed this analysis?
Through discussions with the country staff and collaboration with the conflict analyst, we homed in on leveraging remotely sensed data to understand the environmental drivers of agricultural productivity in the region. Building upon the groundwork laid by the Sudan team’s previous remote sensing analyses, we aimed to extend our understanding beyond observation to predictive insights.
Our primary objective was first to understand the intricate relationships between environmental factors and agricultural outcomes in four states in southern Sudan, a region known as the country’s breadbasket. Second, we desired to develop predictive models capable of forecasting vegetation health and agricultural productivity based on environmental data forecasting, offering valuable guidance for resource distribution and agricultural assistance efforts in the area.
So you crafted the analysis around a decision or a particular set of decisions related to where and how a program team might consider implementing agricultural programming.
That’s right, by focusing our analysis on this specific aspect of the complex system, we hoped to provide practical support to implementing programs that target agricultural resilience in Sudan. This approach allowed us to navigate the vast array of challenges inherent in the humanitarian sector while staying true to the mission of leveraging data-driven insights to effect positive change in conflict-impacted communities.
What do you think are the limitations of using this type of geospatial analyses in a context like Sudan versus, for example, trying to do something similar for agriculture in the US or Europe?
I think geospatial analyses could play a pivotal role in humanitarian efforts and agricultural development, but regional disparities influence their effectiveness in data accessibility, quality, and infrastructure. Collecting comprehensive, ground-based data in low-to-middle-income and conflict-affected areas is beset with challenges, including geographical inaccessibility, logistical constraints, and security risks. These obstacles lead to significant gaps in data, undermining timely and effective responses to crises. Conversely, in high-income countries such as the US and Europe, the agricultural sector benefits from an advanced data infrastructure, where extensive networks of ground stations enrich remotely sensed data. This robust framework supports precise monitoring, planning, and resource management, highlighting a stark contrast in data utilization capabilities between regions. The disparity extends to the availability and granularity of remotely sensed products, which are critical for detailed environmental assessments and agricultural management.
And are there specific ways that disparity in data availability and quality come into play when modeling for humanitarian scenarios?
Definitely. Humanitarian operations in less resource-rich settings often depend on lower-resolution, generic datasets that need more specificity for accurate planning and intervention. In a scenario like Sudan, humanitarian access is already limited and it’s unlikely we can collect ground-based data for calibration and validation. This underscores the challenges of applying geospatial analyses in humanitarian contexts (especially in crisis) and demonstrates a pressing need for enhanced data infrastructure in these vulnerable regions.
When you finished this project, we talked a lot about the risk of people over-interpreting results. Can you talk about that and maybe about how people should try to understand the concept of uncertainty in the forecasts?
In discussing the risks associated with over-interpreting the outcomes of predictive models, it is crucial to underline the challenges that arise with geospatial analyses, especially when conveyed through maps. As a visualization tool, maps are highly accessible yet prone to overinterpretation. Despite their value in decision-making, the clarity of maps can lead users to attribute undue precision and certainty to forecasts, potentially overlooking the inherent uncertainties, the natural variability in environmental data, and complexity of the underlying models.
It would be hard to visualize prediction intervals on a map to convey uncertainty.
That’s right. More specifically, it’s very difficult to convey the intricacy of relationships among variables in a model; plus the need for established environmental thresholds complicates understanding individual drivers and their contributions to forecasts, such as those predicting agricultural productivity. This complexity demands a cautious approach to interpreting map-based outputs so that users remain aware of the simplifications and assumptions embedded within these visual representations.
For this project, you also worked closely with Sarah McArthur, a conflict analyst specializing in humanitarian assistance. How did you find that experience? Do you have any thoughts about how data scientists should be collaborating with humanitarian experts more broadly? How does that play into this conversation on interpretation and uncertainty?
Sarah’s more recent experience in the Sudan context was instrumental for our project, highlighting the essential role of melding specialized context knowledge with data science to tackle complex humanitarian challenges. Acknowledging that we were dealing with an extraordinarily complex system where environmental data alone falls short of explaining or capturing the full scope of challenges, having this technical-contextual partnership enabled us to navigate the complexity effectively. Instead of being overwhelmed by the system’s intricacies, we focused on critical areas where our analysis could make a tangible difference, enriching our quantitative findings with qualitative insights to provide a more holistic understanding of the situation.
I also think that adopting this sort of “triangulation” approach involving integrating diverse data sources, methodologies, expert opinions, and rigorous model validation processes is vital to address these challenges around interpretation. It underscores the power of interdisciplinary collaboration in navigating and making sense of complex humanitarian contexts. By combining rigorous data analysis with deep contextual understanding, we were able to offer insights both informed by data and connected to the realities on the ground. We hope this approach supports the development of interventions that are not only technically actionable but resonant with the needs and conditions of those affected by the conflict in Sudan.
Do you think this analysis can be scaled beyond the borders of Sudan? Do you see any risk or downside in replicating it in other contexts?
When considering the expansion of this analysis from Sudan to other regions I think it is essential to recognize that each area’s unique environmental, socio-economic, and political conditions play a significant role in the application and effectiveness of our methodologies. A strength of our approach in Sudan was our rigorous examination of the connections between environmental variables and agricultural productivity, including the impact of seasonal changes and lagged variables.
Conducting a comprehensive model exploration phase is critical for the analysis to be successfully adapted to new locations. This involves adjusting and testing the model to ensure it fits the new context while also considering challenges such as inaccuracies from cloud cover in remotely sensed data. However, beyond the technical adjustments, an invaluable aspect of this process involves working closely with the country team on the ground. Their insights into the daily realities, challenges, and opportunities within their specific locales are indispensable for tailoring the analysis to be as relevant and impactful as possible.
The potential benefits of replicating this analytical approach in other regions are significant. As a predictive and forward-looking geospatial tool, it can enhance strategic resource allocation, complement other remote sensing tools available, and support informed decision-making to optimize agricultural productivity.
What do you think the next steps are for this project? Were there any facets of food security, agriculture, and conflict that you wanted to explore that you did not have the opportunity to look at this time around?
Given the many pathways and opportunities that exist with this analysis, my interests in this project could well encompass a full-time PhD dissertation. A key focus area is testing the integration of socio-economic, conflict data, and additional environmental variables into our model. This integration would be crucial for understanding the multifaceted influences on agricultural productivity, which could lead to more nuanced interventions, programs, and support.
Another vital area involves refining our predictive models through exploring alternative machine learning techniques and methodologies, and tracking the forecasting performance of the model as time passes. This refinement aims to improve the accuracy and applicability of our forecasts, which is essential for making well-informed decisions. Expanding the study geographically within Sudan and beyond will allow us to assess the model’s versatility and pinpoint general and specific factors affecting agricultural productivity across different settings.
Thanks, Nick. These are really comprehensive answers! The analysis done on this round of our work with Sudan has really pushed us forward in terms of how we engage with remote sensing data and opportunities for using AI and ML technologies. We’re really looking forward to the feedback we receive on the reports from both our program teams as well as the broader community of NGOs doing work in Sudan.
Interested in our work? Reach out at dataforimpact@mercycorps.org.
